Gauransh Sharma

Remote · betterstory.cc

• Built LLM character DMs end-to-end: a chat spec authored in the admin dashboard, Claude Sonnet with structured outputs, reply chips and reactions, message reporting, and cron-driven jobs that generate and send proactive character-initiated openers.
• Built audiobook narration on ElevenLabs: per-chapter and whole-story generation with paragraph timings, an admin console to generate and audition chapters, and an Expo player with background playback, lock-screen controls, scrubbing, and resume.
• Built offline reading in Expo: a new API prefetch flag (no progress side effects) feeds a 5-chapter lookahead cache, and an outbox queues choices and completions and replays them exactly once on reconnect.
• Also built in-app story search (tiered ranking over titles and tags, admin-managed trending terms), image support in push notifications (admin picker, server-side framing for Android, iOS rendering), and GA4 conversion events for Google Ads.

Remote · oasishealth.app

• Contributed to a new evidence-based product scoring system and built its mobile experience in React Native, covering onboarding, personalization, and product detail surfaces, verified across iOS and Android.
• Shipped a persisted scoring pipeline with a database migration and batch scorer so 500k+ products could be ranked server-side via indexed lookup, replacing per-request computation that took 0.5-2s per item.
• Delivered a contaminant-scoring correctness fix (speciation handling for arsenic and mercury) and hardened the lab-results ingestion adapter to recover measurements that were previously missed.
• Rebuilt admin panel surfaces and contributed to the brand submission portal, covering table UX, filtering, and review workflows used by the internal operations team.

USA · Remote · prana.health

• Designed and shipped a unified LLM coaching agent on AWS Bedrock (Claude Sonnet) with LangChain tool-calling, SSE streaming, a thread-scoped tool cache, and multi-stage pipelines for onboarding, DEXA scan processing, and blood biomarker extraction via AWS Textract.
• Built and shipped an MCP server exposing the full coaching toolset (food and medication logging, workout tracking, DEXA and biomarker insights, personalized recommendations), letting users run the entire Prana AI chat experience from Claude or ChatGPT through a connector.
• Implemented a hybrid RAG pipeline using Pinecone (BM25 + Titan Embeddings v2) and Mem0 for persistent cross-session memory, enabling personalized coaching context that compounds over time.
• Architected a large-scale FastAPI + PostgreSQL backend with dual-auth (Supabase JWT + SHA256-hashed MCP API keys), row-locking transaction helpers, Supabase RLS policies, and a custom in-memory rate limiter.
• Engineered a production Flutter client with real-time AI chat (SSE streaming with embedded charts), barcode and image scanning, food and medication logging, ElevenLabs voice integration, and offline-first persistence with background sync.

Turkey · Remote

• Designed a two-tier matching system combining geospatial indexing with a weighted compatibility scoring engine, exposed via optimized REST APIs and consumed in a React Native client.
• Optimized React Native state management to achieve smooth 60fps gesture-driven animations; built and maintained GitHub Actions CI/CD pipelines enforcing test coverage and API contract validation.

Sweden · Remote · edventures.ai

• Defined the mobile architecture roadmap and led migration to a serverless backend; mentored the team on Flutter Clean Architecture, reducing technical debt through structured code reviews.

Breaks a goal into a graph of tasks and runs each one with a coding agent (Claude Code, Codex, Gemini CLI, or a model on NVIDIA NIM) in its own git worktree. A task merges only when its checks pass, an optional reviewer agent approves, and the checks pass again on the merged tree; a merge that breaks the tree is rolled back, and conflicts go to a narrow resolver agent that keeps both sides. Work lands on an integration branch, never your own, and an interrupted run resumes where it stopped. Ships as a CLI, a live dashboard, and a Tauri desktop app for macOS and Windows.

TypeScript, Node.js, Git worktrees, Tauri, Claude Code, Codex, Gemini CLI

WebsiteSource

Send a meeting link to a Telegram bot and Tekki shows up in the call as its own participant, listens, thinks, and answers out loud with a face that reacts. Google Meet has no bot API, so it becomes the devices instead: getUserMedia is replaced inside the page and handed a canvas for video and a WebAudio node for audio. Hearing works the other way round — Chrome never returns decoded WebRTC audio to WebAudio, so the container records its own PulseAudio sink. Every model runs locally; nothing about the conversation leaves the machine.

Playwright, WebRTC, Docker, Ollama, whisper.cpp, Kokoro TTS, Node.js, Canvas2D

WebsiteSource

Autonomous reasoning engine on LangGraph with a recursive Generator-Critic loop, cutting hallucinations through self-correcting multi-agent feedback. Lifelong-learning memory system built on ChromaDB (Vector RAG) and DSPy BootstrapFewShot enables runtime self-improvement without retraining, with a human-in-the-loop Chainlit interface where user critiques propagate directly into vector memory.

LangGraph, DSPy, Llama-3, ChromaDB, RAG, Chainlit

Source

Chat application with AI-driven sentiment analysis to suggest replies and analyze user emotions. Built a responsive cross-platform interface using Flutter with real-time messaging.

Flutter, Firebase, React, AI, Sentiment Analysis

WebsitePaper

User/Order microservices with real-time monitoring, implementing fault tolerance via Circuit Breaker and retry mechanisms. Utilized Docker for containerization with Fluentd/Prometheus for log collection and chaos testing.

Python, Flask, Docker, Prometheus, Chaos Engineering

Source

Minimalist typing speed test built from scratch with vanilla JavaScript. Supports three test modes (words, timed, and quote) with a floating real-time caret, live WPM and accuracy tracking, and a post-test performance chart rendered on the Canvas API with no charting libraries. Features a 1,500-word pool, per-keystroke error display, 30-result local history, and full keyboard navigation.

JavaScript, Canvas API, Firebase, Vanilla JS

Website

Go CLI that reads your staged git diff and generates a conventional commit message using Kimi K2 via NVIDIA NIM. Zero config beyond an API key. Installable via go install or Homebrew.

Go, Kimi K2, NVIDIA NIM, CLI

Source